Hi Privacy Focused People,

New updates have arrived for Note67

Download at https://note67.com/download

What’s New

v0.1.25 · July 2026

Last month was about making transcripts readable. This one is about the edges — the short replies, the crosstalk, and the odd artifacts that speech-to-text leaves behind. Individually they’re small. Together they’re the difference between a transcript you trust and one you skim past.

Short replies count as speech

Live transcription works in windows, and a reply short enough to sit inside one — “hello,” “thanks,” “yep, agreed” — could fall through the gaps without ever landing on a line. Meetings are mostly made of these. A decision usually isn’t a paragraph; it’s someone saying “fine, let’s do that.” They’re now captured like anything else.

Talking over each other is normal

Echo deduplication exists because a microphone picks up your own speakers when you’re not wearing headphones, and you end up reading the same sentence twice. The old logic was too eager: if a mic segment overlapped in time with a system audio segment, it was treated as echo and dropped.

But overlapping in time is just what a conversation sounds like. Interruptions, agreement, someone jumping in to correct a number — all of it was being deleted as duplicate audio. A mic segment is now only considered echo when it actually repeats the words from an overlapping system segment. Timing alone isn’t enough anymore.

No more stage directions

Whisper occasionally editorializes. Rather than transcribing speech, it describes it — *sighs*, *Slow's voice*, *door closes* — and those descriptions were being written into transcripts as though someone had said them out loud. They’re now filtered out before they reach the note.

Live transcript stays where you’re looking

When you continue recording on an existing note, new lines were being inserted above the earlier ones, which meant the transcript reordered itself while you were reading it. New text now appends below what’s already there, and the view follows along as it arrives instead of leaving you to chase it.

Turbo Q8 is the new recommendation

The recommended transcription model is now the Q8 quantized build of large-v3-turbo. Same quality, same speed, roughly half the download and less memory while it runs. If you’re setting up a new machine, this is the one the first-run wizard will point you at. Existing models keep working — nothing is being retired.

Fixes under the hood

• Playback after continuing a recording no longer comes out silent.
• Retranscription is no longer destructive. If it’s interrupted partway through — you quit, the app crashes, the machine sleeps — your existing transcript survives instead of being left half-replaced.
• Startup is faster and the app bundle is smaller.

As always: everything above happens on your device. No audio leaves your machine, and there’s no account to create.